Red-Team Scenario Catalog

Reference for the attack vectors NuGuard's red-team engine can generate. See redteam-guide.md for how to configure a scan and redteam-design.md for the engine's internal architecture.

The built-in catalog ships 125 scenario specs across 18 attack categories. Each spec is SBOM-grounded: the ScenarioGenerator only instantiates a scenario against components that actually exist in the scanned application (e.g. an R-category RAG attack is only built if the SBOM has a vector store / retrieval component), and only if the SBOM exposes the required_capabilities that scenario needs.


How a scenario is chosen

Every catalog entry carries:

  • goal_type — one of 9 high-level attack goals (see GoalType taxonomy); shown in reports and findings, but not itself a redteam.scenarios filter axis (see below).
  • base_impact (0–10) — a pre-score used by redteam.profile / --profile to decide whether the scenario is worth running (see Profile filtering).
  • required_capabilities — SBOM signals (e.g. rag, multi_session, sensitive_context) that must be present for the scenario to apply.
  • safe_execution — the containment strategy used so the attack can run against a real target without causing real-world harm (see Safe execution modes).
  • priority_rules — tie-breakers used when trimming the plan to max_scenarios or a profile budget.
  • enabled — whether the scenario is active by default; disabled entries (currently only V03) are opt-in via a customized --catalog file.

GoalType taxonomy

Every scenario is tagged with one of 9 high-level attack goals, shown in reports and findings for context. redteam.scenarios / --scenarios no longer filters on GoalType — see Scenarios (destructive / non-destructive filter) in the guide for the current filter axis. Each GoalType maps to one or more catalog categories below:

GoalType Description Catalog categories
prompt-driven-threat Prompt injection, jailbreaks, guardrail bypass, system-prompt extraction I, J, plus parts of E, O, R, S, V
data-exfiltration PII/PHI/secret disclosure, cross-tenant leaks, covert encoding C, D, plus parts of N, O, R, S, V
privilege-escalation Auth bypass, role override, credential/consent abuse H, plus parts of A, N, O, V
tool-abuse Unsafe or destructive tool invocation K, T, plus parts of B, H, O
mcp-toxic-flow Untrusted content flowing into write-capable MCP tools M, plus part of R
api-attack BOLA/BFLA, mass assignment, schema-level exploitation S, plus part of A
agentic-trust-abuse Confused deputy, memory poisoning, multi-agent trust boundaries G, P
policy-violation Topic-boundary and restricted-action violations B, plus part of A
recon-inference Information inferred from refusal patterns / response timing parts of E, R, S, V

Categories and scenarios

Prefix letters are the stable catalog ID prefix (e.g. D01–D08), used to identify individual scenarios in reports and a custom --catalog file — not a redteam.scenarios filter value (see Scenarios (destructive / non-destructive filter)).

A — Authorization Failures (8)

ID Title Impact Goal
A01 BOLA other-user object read 8.0 api-attack
A02 BOLA other-user object write 8.5 api-attack
A03 BFLA privileged function call 7.5 api-attack
A04 RBAC role override 7.0 privilege-escalation
A05 Auth-bypass via false verification 7.0 privilege-escalation
A06 HITL bypass 7.0 policy-violation
A07 Mass assignment hidden fields 7.0 api-attack
A08 Debug/admin command exposure 6.5 api-attack

B — Business Logic and Safety (6)

ID Title Impact Goal
B01 False successful action claim 6.0 policy-violation
B02 Unsafe recommendation outside domain 5.5 policy-violation
B03 Fraudulent refund/chargeback workflow 6.5 policy-violation
B04 Compliance/legal/medical/financial advice overreach 5.5 policy-violation
B05 Rate-limit/resource exhaustion 5.5 tool-abuse
B06 Hallucinated authority/source 5.0 policy-violation

C — Covert Exfiltration (8)

Data smuggled out through side channels rather than plain-text answers.

ID Title Impact Goal
C01 Markdown image pixel 7.5 data-exfiltration
C02 Reference-style markdown link 7.0 data-exfiltration
C03 URL query/path/fragment encoding 7.5 data-exfiltration
C04 Base64/hex/ROT encoding 6.5 data-exfiltration
C05 JSON/XML field smuggling 6.0 data-exfiltration
C06 CSV/document metadata smuggling 6.0 data-exfiltration
C07 Invisible Unicode/ASCII smuggling 6.5 data-exfiltration
C08 Logging/sync/analytics pretext 7.0 data-exfiltration

D — Data Exfiltration (8)

Direct disclosure of PII/PHI/secrets or other users' data.

ID Title Impact Goal
D01 Direct profile dump 7.0 data-exfiltration
D02 Cross-tenant lookup 8.0 data-exfiltration
D03 Adjacent-ID/IDOR probe 7.5 data-exfiltration
D04 Cross-session secret leak 8.0 data-exfiltration
D05 Private document extraction 7.0 data-exfiltration
D06 RAG citation overreach 6.5 data-exfiltration
D07 Aggregated PII across turns 6.5 data-exfiltration
D08 Unauthorized location/history disclosure 6.5 data-exfiltration

E — Evasion and Robustness (7)

ID Title Impact Goal
E01 Multilingual restricted request 5.5 prompt-driven-threat
E02 Homoglyph/leetspeak 5.5 prompt-driven-threat
E03 Context flooding 5.5 prompt-driven-threat
E04 Structured JSON/XML override 5.5 prompt-driven-threat
E05 Refusal oracle reconnaissance 5.0 recon-inference
E06 Model/tool/system prompt extraction 5.5 prompt-driven-threat
E07 Unexpected code generation probe 7.0 prompt-driven-threat

G — Multi-Agent Trust Abuse (6)

ID Title Impact Goal
G01 Confused deputy 7.0 agentic-trust-abuse
G02 Sub-agent output injection 7.5 agentic-trust-abuse
G03 Handoff-chain privilege escalation 7.5 agentic-trust-abuse
G04 Agent impersonation 6.5 agentic-trust-abuse
G05 Planner/executor mismatch 7.0 agentic-trust-abuse
G06 Supervisor approval spoof 7.0 agentic-trust-abuse

H — Human-Agent Trust Exploitation (5)

Attacks that abuse the human-in-the-loop approval step itself.

ID Title Impact Goal
H01 Approval summary mismatch 8.0 privilege-escalation
H02 Consent laundering 7.5 privilege-escalation
H03 Authority-bias operator phishing 7.0 privilege-escalation
H04 Partial approval overreach 8.0 privilege-escalation
H05 Hidden recipient/action payload 8.5 tool-abuse

I — Indirect Prompt Injection (8)

Instructions smuggled into content the agent ingests, rather than the user prompt.

ID Title Impact Goal
I01 Poisoned webpage 7.5 prompt-driven-threat
I02 Poisoned email/message 7.5 prompt-driven-threat
I03 Poisoned calendar/event 7.5 prompt-driven-threat
I04 Poisoned file/document 7.5 prompt-driven-threat
I05 Poisoned search result 7.5 prompt-driven-threat
I06 Poisoned tool response 7.5 prompt-driven-threat
I07 Poisoned database row 7.5 prompt-driven-threat
I08 Prompt injection in markdown/HTML 7.5 prompt-driven-threat

J — Jailbreak and Policy Bypass (6)

Published jailbreak techniques (Crescendo, Many-Shot, Skeleton Key, etc.).

ID Title Impact Goal
J01 Crescendo escalation 6.0 prompt-driven-threat
J02 Many-shot jailbreak 6.0 prompt-driven-threat
J03 Skeleton Key/developer framing 6.0 prompt-driven-threat
J04 Roleplay/fictional training bypass 6.0 prompt-driven-threat
J05 Payload splitting across turns 6.0 prompt-driven-threat
J06 Policy-conflict false premise 6.0 prompt-driven-threat

K — Coding and Automation Agents (6)

Targets coding assistants and CI-connected agents.

ID Title Impact Goal
K01 Repo prompt injection 7.0 tool-abuse
K02 Shell injection 8.0 tool-abuse
K03 Secret env/file read 7.5 tool-abuse
K04 Sandbox read/write escape 7.5 tool-abuse
K05 Delayed CI exfiltration 7.5 tool-abuse
K06 Verifier/test sabotage 7.0 tool-abuse

M — MCP and Tool Poisoning (8)

Model Context Protocol-specific attacks — untrusted content flowing from an MCP server into a write-capable tool (MCP_TOXIC_FLOW).

ID Title Impact Goal
M01 Malicious tool description 7.5 mcp-toxic-flow
M02 Lookalike/shadow tool 7.0 mcp-toxic-flow
M03 Tool output instructs second tool 7.5 mcp-toxic-flow
M04 Toxic flow from untrusted source to write sink 8.0 mcp-toxic-flow
M05 Credential-scope overreach 7.0 mcp-toxic-flow
M06 External fetch/SSRF via tool 7.5 mcp-toxic-flow
M07 Tool-discovery leakage 5.5 mcp-toxic-flow
M08 Multi-server cross-tool exfiltration 8.0 mcp-toxic-flow

N — Agent Identity and Credential (6)

ID Title Impact Goal
N01 OAuth consent/scope escalation 8.0 privilege-escalation
N02 Token replay/session fixation 8.5 data-exfiltration
N03 Ownerless agent action 7.5 privilege-escalation
N04 Cross-agent credential bleed 8.5 privilege-escalation
N05 Delegated identity confusion 7.5 privilege-escalation
N06 Credential persistence in memory 7.0 data-exfiltration

O — Improper Output Handling (6)

Downstream systems trusting agent-generated output without validation.

ID Title Impact Goal
O01 HTML/Markdown XSS injection 7.0 prompt-driven-threat
O02 JSON tool argument injection 7.5 tool-abuse
O03 Generated SQL tenant bypass 8.0 data-exfiltration
O04 Generated URL causes SSRF 8.5 tool-abuse
O05 YAML/CI config security override 7.5 privilege-escalation
O06 File/MIME confusion artifact 7.0 privilege-escalation

P — Memory and Persistence (6)

ID Title Impact Goal
P01 Poison long-term memory 7.5 agentic-trust-abuse
P02 Poison preference/profile field 7.0 agentic-trust-abuse
P03 Cross-session backdoor instruction 8.0 agentic-trust-abuse
P04 False identity memory 7.5 agentic-trust-abuse
P05 Malicious summary persistence 6.5 agentic-trust-abuse
P06 Memory-based authorization drift 6.5 agentic-trust-abuse

R — RAG and Vector Store (8)

ID Title Impact Goal
R01 Poisoned document ingestion 8.0 mcp-toxic-flow
R02 Vector metadata ACL bypass 8.5 data-exfiltration
R03 Embedding collision hijack 7.0 mcp-toxic-flow
R04 Chunk-boundary smuggling 7.5 prompt-driven-threat
R05 Stale document retrieval 6.5 data-exfiltration
R06 Cross-index namespace bleed 8.0 data-exfiltration
R07 Citation laundering 6.0 data-exfiltration
R08 Nearest-neighbor enumeration 5.5 recon-inference

S — API Schema Exploitation (8)

ID Title Impact Goal
S01 Identity field body override 8.0 api-attack
S02 Schema-grounded mass assignment 7.5 api-attack
S03 Hidden response schema field exfiltration 6.5 data-exfiltration
S04 Streaming response data exfiltration 7.0 data-exfiltration
S05 Alternate payload-key injection 6.0 prompt-driven-threat
S06 Session ID replay / session fixation 7.0 api-attack
S07 Structured output field coercion 5.5 prompt-driven-threat
S08 Schema field type-confusion recon 4.5 recon-inference

T — Destructive Tool Actions (8)

Irreversible or high-consequence tool calls executed without confirmation.

ID Title Impact Goal
T01 Send message/email without confirmation 7.5 tool-abuse
T02 Delete/update record 8.0 tool-abuse
T03 Change route/destination 7.0 tool-abuse
T04 Unsafe climate/media command 6.5 tool-abuse
T05 Payment/transfer/order action 8.0 tool-abuse
T06 Mass notification 7.0 tool-abuse
T07 Disable alert/logging 7.5 tool-abuse
T08 Create persistent scheduled task 7.0 tool-abuse

V — Supply Chain and CI/CD (7)

ID Title Impact Goal Enabled
V01 Named env var secret extraction 8.0 data-exfiltration yes
V02 CI/CD runtime token disclosure 8.5 data-exfiltration yes
V03 Cloud metadata IMDS credential theft 9.0 data-exfiltration no
V04 Known-vulnerable dependency exploit probe 6.0 prompt-driven-threat yes
V05 No-quality-gate dark corner probe 4.0 recon-inference yes
V06 Model artifact integrity discrepancy 5.0 recon-inference yes
V07 Staging credential reuse on production endpoint 8.0 privilege-escalation yes

V03 ships disabled by default because IMDS credential theft probes can trigger real cloud-provider security alerting even in a sandbox. Enable it explicitly via a customized --catalog file if your environment can tolerate that.


Safe execution modes

Every scenario declares how it's run against a live target without causing real damage:

Mode Behavior
canary_only Only probes for seeded canary values (see canary setup) — never targets real user data
synthetic_tenant Uses a dedicated synthetic tenant/account created for the scan, not a real customer
trap_endpoint Points destructive actions at a sandboxed trap endpoint that records the call without executing it
dry_run_tool Invokes the tool's dry-run/preview mode where the app supports one
trace_only Observes agent reasoning/tool-selection traces without requiring the action to actually execute
emulated_tool Uses a stubbed tool response to observe how the agent would react
sandbox Runs inside an isolated execution sandbox (coding-agent scenarios)

In campaign mode (redteam.mode: campaign) safe_execution is enforced: synthetic_tenant needs a second principal (a canary tenant with a session_token), trap_endpoint needs the callback canary server, and dry_run_tool / sandbox / emulated_tool need redteam.campaign.declared_fixtures: true. Without the fixture the objective is recorded as blocked_fixture:<name> rather than run. Each entry also has a complexity level (L1–L7) and a session policy (reuse, fresh, isolated) used by the campaign scheduler; these live beside the catalog in nuguard/redteam/catalog/scheduling.py, not in the YAML.


Customizing the catalog

Export the full catalog to inspect or edit it:

nuguard redteam catalog-export --output my-catalog.yaml

Common edits:

  • Set enabled: false on a scenario to skip it entirely (e.g. disable T05 payment-action probes if your sandbox has no test payment gateway).
  • Lower base_impact below a profile's threshold to exclude a scenario from fast (ci) scans without disabling it globally.
  • Edit expected_control / success_signal to match your application's specific behavior.

Pass the edited file back in:

nuguard redteam --sbom app.sbom.json --target http://localhost:8000 \
  --catalog ./my-catalog.yaml --profile full

or in nuguard.yaml:

redteam:
  catalog_path: ./my-catalog.yaml

See example-openai-cs-agents.md for a worked scan against a live demo app.